Deep reinforcement learning method for autonomous exploration of unknown underground zones using aerial robots with 3D LIDAR

Abstract: The capability of robots to works in dangerous environments like underground mines can allow workers to work in a safer environment. In particular mapping and explore mines and tunnels can save lives and money. For this kind of missions, UAVs are often chosen, thanks to their high mobility in 3D spaces. However, to do that, the robot needs to work in a complex environment with reduced visibility and poor connections. Artificial intelligence and in particular Deep Reinforcement Learning (DRL) is becoming a more popular approach to solve this kind of problems, but the implementation of this techniques is still difficult and limited, especially when treating with 3D data. For these reasons, the aim of this work is to provide a new approach to solve the problem of autonomous exploration of 3D underground environments using DRL. We present and explain the approaches which has worked and why other approaches have not, with particular attention in the resources used. Moreover, we implement the new technique paying attention to provide a system which can be easily transferred in a real UAV. Hence, we test the new approach with the actual state of the art in the field, explaining which technique use today and why DRL is a valuable alternative to more analytical approaches used nowadays.